Ask a plain-English question about New York law. Get an answer grounded in real opinions and statutes — every claim anchored to a source you can open and read for yourself.
Ask the corpus from the panel on the right — or open the full page. Research tool, not legal advice; supervised by Derek Soltis, Esq. (NY Bar).
Live · pulled from Aurora via /stats on page load
Lex.NY answers plain-English questions about New York law with the receipts attached — every claim anchored inline to a real opinion or statute you can click straight through to. Search by issue, check a brief’s citations, or explore judge analytics, all drawn from 1.32M NY opinions and the full statutes in AWS Aurora.
It retrieves real documents first and abstains when the corpus doesn’t cover you — so it can’t invent a case.
Try it yourself →1.32 million indexed opinions — the NY Court of Appeals, all four Appellate Divisions, and the federal SDNY / EDNY — plus all 137 NY Consolidated Laws and ~40,000 statute sections. Embedded with mxbai-embed-large (1024-d) into pgvector on AWS Aurora.
Every opinion's CITES and APPLIES edges live as relational tables in Aurora and are traversed with recursive CTEs — most-cited decisions, judge influence, good-law signals — all in the same engine as the vectors.
The drafting model works under a system prompt that forbids any unsourced claim. Every paragraph anchors to a numbered marker pointing into the retrieved context. If the corpus doesn't cover the question, Lex.NY abstains.
pgvector ANN, the relational citation graph, and full-text search in a single database. IAM-authenticated — passwordless, with a fresh signed token per connection.
Every answer is drafted under a no-unsourced-claims system prompt. The model only sees what was retrieved from the corpus first — no outside knowledge bleeds in.
The shared vector space for 3.5M indexed passages. Query and corpus are embedded identically, so retrieval matches on meaning, not keywords.
Sub-second vector search across the embeddings table, tuned for recall on NY-specific retrieval and pre-warmed into memory for the demo.
Most-cited decisions, judge influence, and case-to-case treatment — computed by traversing CITES / APPLIES edges, no separate graph store.
Serverless functions colocated with Aurora for low-latency retrieval. Streaming answers over Server-Sent Events, deployed from GitHub.
Built on AWS · Aurora PostgreSQL Serverless v2 · Deployed on Vercel
Free, public, rate-limited, attorney-supervised. The model can’t see anything that wasn’t retrieved from the corpus first.